LLPROBeta

Home / Agents / Change Witness

Change Witness

A signed, dated receipt that a team stated a passing staged check for a change before it reached a shared layer, with a lookup by fingerprint later.

Live

What it does
A signed, dated receipt that a team stated a passing staged check for a change before it reached a shared layer, with a lookup by fingerprint later.
Who it is for
Agents and their builders.
Start in one step
Use it on this page, or ask your agent to call change_witness at llpro.ai/mcp.
What to expect
Live: open today. Free: 20 records a day per caller. Paid records (0.02 USD) not charged yet.

New to this? Start here.

What it does

Change Witness gives a release team, or its agent, a signed and dated receipt for a stated passing staged check, taken before the change goes out to a layer many others share (a config store, an edge network, a package registry, a model or prompt rollout). Send the change fingerprint (the sha256 of the release artifact or config), the check you ran (its name, result pass and the sha256 of its log) and the layer it is about to reach. LLPRO signs that it received those values at that time and keeps the record; later anyone with the fingerprint can look it up and check each signature with the public key below. It never runs the check, sees the code or says the change is safe, and it cannot see when the change went out: keep the record with your own deploy time to show the check came first. What a record says: A Change Witness record says that LLPRO received, at received_at (UTC, LLPRO's clock), four values: a fingerprint (the sha256 of a release artifact or config), a staged check the sender says passed (its name, result pass, and the sha256 of its log) and the name of the shared layer the sender says the change is about to reach. Hashes are stored as lowercase hex with any 'sha256:' prefix and surrounding spaces removed, and a result of pass or passed in any letter case is stored as pass; names are stored exactly as sent. Only a passing result is recorded. The same four values sent again return the first record unchanged (replayed true), so a record's time is the first time LLPRO received them within the last 400 days. A lookup by fingerprint returns the stored records for it, oldest first, 20 at a time, with the total, and a signed lookup observation that lists the sha256 of each returned record (the sha256 of its observation in canonical JSON), also when there are none. Records are kept 400 days from received_at, then deleted.

Who it is for
Release agents and the teams behind them that push updates to a shared layer and want a dated outside receipt for the staged check they ran, for a post-release review, an audit trail or a customer who asks.
Price
Free, no sign-up: 20 records a day per caller; lookups and replays are free. Paid records (0.02 USD a new record past 20) are not switched on yet: until they are, records past 20 are not charged and run free within the limits below. The opt-in test-money mode of the campus paid tier today asks a test payment for every call of a tool, lookups included; lookups and replays will not be billed when paid records switch on.
Limits
Only a passing check is recorded (result pass); any other result is refused and nothing is stored. Fingerprint and log hash: 64 hex characters each (sha256; a 'sha256:' prefix is fine). Check name and layer: 1 to 80 characters of letters, digits, spaces and . _ : / + # = -, no links and no 'llpro'; anyone who knows the fingerprint can read them, so never put a secret or personal data in them. Arguments up to 2048 bytes. Calls an hour: 300 per caller address (an IPv6 /64 counts as one) and 1200 per network (IPv4 /24, IPv6 /48). New records a day: 200 per address and 400 per network; and at most 6000 new records an hour for all callers together. Sending the same four values again returns the first record and does not count as new. A lookup returns 20 records per answer, oldest first, with the total; use offset for the next 20. The store holds up to 2,000,000 records. Kept by the tool: the fingerprint and each signed record with its signature (check name, result, log hash and layer inside it), for 400 days; and, for 2 days, a keyed hash of your address and of your network to count the daily limit. Not kept: your address itself, the artifact, the config or the log. The campus gateway separately keeps call metadata with a hashed caller id and the call time for 90 days, and through the agent-to-agent door the task answer for 7 days; those could be matched to a record's time. Records are deleted 400 days after they were made; the nightly backup copy of the store is kept up to 7 more days. The gateway's call metadata (tool, status, time, a hashed caller id, country, user agent family, referring site) is used for traffic counts and abuse checks, never the request contents.

For developers: try it

POST https://llpro.ai/mcp
{"jsonrpc":"2.0","id":1,"method":"tools/call","params":{"name":"change_witness","arguments":{"fingerprint":"<sha256 of the artifact>","check":{"name":"staging-smoke","result":"pass","log_sha256":"<sha256 of the check log>"},"layer":"prod-config"}}}
-> replayed (false the first time), record_sha256, and a signed observation: received_at, fingerprint, check, layer, rule_sha256, signer and the does_not text. Later: {"action":"lookup","fingerprint":"<sha256 of the artifact>"} -> total, next_offset, the stored records (20 per answer, each with its signature) and a signed lookup observation, also when there are none. The same arguments work on REST: POST https://llpro.ai/api/v1/tools/change_witness with the arguments as the JSON body.

For agents

Tool change_witness on https://llpro.ai/mcp and in the tools list. Signed answers are checked with the public key below (signer llpro1:fwnadwdck6ckzlfqoewshdiqie): an Ed25519 signature over the observation as canonical JSON (sorted keys, no spaces).

Public key (base64): jeBQa747S/UsyFkT3NeRI0M4f0z9D0K6WjMsZxkEcpM=

Listed in MCP registries and served by the Campus gateway (MCP, agent-to-agent and REST doors). Call it after the staged check passes and before the change reaches the shared layer; keep the whole answer with the release. The signature covers the observation, including the signer id and the does_not text. Canonical JSON means sorted keys, no spaces, UTF-8 and no ASCII escaping (in Python: json.dumps(observation, sort_keys=True, separators=(',', ':'), ensure_ascii=False)). Errors come back as isError with a plain reason: a fingerprint or log hash that is not 64 hex characters, a result other than pass (422, nothing stored), a name or layer with other characters or over 80, an unknown field, or a limit reached (429).

Next

For agents: this page as Markdown · JSON · llms.txt · MCP endpoint llpro.ai/mcp · agent card